A Novel Method for Classification of Exudates and Retinal Sheen in Fundus Photographs for Automated Disease Grading

Abstract

Purpose : Bright structures such as retinal sheen that are especially common in young population, due to the extreme transparency of the eye lens and vitreous humor, mislead exudates detection by automated methods. This paper aims at developing novel feature descriptor and applying machine learning technique to differentiate sheen from exudates.

Methods : The image is first resized to have optic disk (OD) diameter of 200 pixels. We computed reproducible orientation of the image,θ=tan^-1((y_M-y_OD)/(x_M-x_OD)),where (x_M,y_M) and (x_OD,y_OD) are the coordinates of the macula center and OD center respectively; which is then used to rotate the image. A local descriptor is then computed for each of the points (i.e. pixels) identified as bright lesions (e.g. exudates, sheen). A square region of size 36×36 around the point is considered. This region is further slipt up into smaller 4×4 square sub-regions. For each sub-region, we compute Haar wavelet responses, d_x,d_y,d_xy at 3×3 regularly spaced intervals using a 4×4 window. The responses are then weighted with a Gaussian of σ=12 centered at the point. The wavelet responses and their absolute values are summed up over each subregion and a 6-dimensional vector v = (Σd_x,Σd_y ,Σd_xy,Σ|d_x|,Σ|d_y|, Σ|d_xy|) is formed. Vectors computed over all the sub-regions are then concatenated to form the descriptor of length 96 to represent the point. The wavelet responses are invariant to bias in illumination. The descriptor is finally normalized to have unit length and invariance to contrast is achieved. A Random Forest classifier is applied to classify each pixel into exudates or sheen based on its description.

DIARETDB1, e-ophtha-EX, and in-house datasets were used for the experiments. An experienced grader outlined the exudates and sheen on the images. Images from all the datasets were finally merged. Table 1 summarizes the image datasets in detail.
Ten folds cross-validation was performed.

Results : The overall classification accuracy was 99.2% (99.6% for exudates and 99.01% for sheen). The precision and recall for exudates were 0.998 and 1.0, and for sheen these values were 1.0 and 0.991.

Conclusions : A novel method to differentiate sheen from exudates is proposed. Experiments show that sheen can be differentiated from exudates by the proposed method with an accuracy of 99%.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Figure 1: (a) Haar wavelets (b) Example sheen